China AI vs. USA AI
Nine days ago we said the model layer was commoditising faster than people thought, and that the thing worth watching was what happens when these models get geo-blocked. Both halves of that are now visible in the data. The interesting part is that the door is being closed from both ends.
@toptraders0x · @alliekmiller · @TechBuzzChina · @Mayhem4Markets · @demian_ai · @jun_song · @glocalinvestor · Booz Allen · Axevil Capital
The number that ended the argument
For two years the question “are Chinese models actually good?” was answered with benchmarks, and benchmarks are easy to argue with. The answer arrived instead as a routing decision. On OpenRouter — the aggregator where a developer picks a model on price and quality rather than brand — Chinese open-weight models went from a rounding error to nearly half of everything.
Two different denominators, deliberately not merged. Green: Chinese open models as a share of all OpenRouter traffic (Axevil Capital). Blue: share of tokens routed by US companies, which OpenRouter data via CNBC put above 30% every week since 8 Feb 2026, peaking at 46%, against an 11% average the year before and 4.5% in H1 2025.
Two caveats you should hold onto, because most people quoting this number drop them. First, the denominators differ: some cuts measure all OpenRouter traffic, others only the tokens routed by US companies, and they are not the same series. Second, OpenRouter is not the market — it is the part of the market that shops. Enterprise contracts signed directly with OpenAI and Anthropic never touch it. What it measures precisely is the segment where switching is cheapest, which is exactly why it moves first.
And it moved. Xiaomi now processes more tokens on OpenRouter than OpenAI does. Anthropic sits at 15.3%, OpenAI at 7.4%. Six Chinese models rank above Claude by volume.
What the frontier premium is actually buying
The cost side is where the argument stops being about ideology. Here is every number we could source, plotted honestly.
GPT-5.5 costs 7× what Kimi does for the identical score. The one real gap is Opus: 7.1 points over the best open model, at 5× the price. Source: Axevil Capital, compiling SWE-Bench Pro and published API pricing.
Read the shape, not the dots. GPT-5.5 charges 7× what Kimi charges for the identical SWE-Bench Pro score. That is not a premium for quality; it is a premium for procurement comfort. The one defensible price in that chart is Opus: 7.1 points clear of the best open model, and that gap is real work on hard problems.
Which is the whole thesis in one line. If your workload is the hard 10% — the debugging session that has to converge, the analysis that cannot be re-run — you buy the frontier and the price is fine. If your workload is the routine 90% that agents now generate in bulk, you are paying 7× for a rounding error, and every quarter you keep doing it the finance team notices harder.
They have noticed. Coinbase cut its internal AI spend by nearly half by moving engineers onto GLM-5.2 and Kimi. Palantir’s Alex Karp has said some US government customers moved off proprietary models onto open ones. Lindy moved its entire traffic. This is not developers playing with a cheap toy; it is a CFO line item.
The two sides
The remaining moat of US frontier labs is narrative, not capability. The story that Chinese models are just distilled knock-offs with backdoors has outlived the evidence, and buyers have started voting with routing tables.
@jun_songThe “good enough” moment is arriving the way it always does. Once a model answers 95–99% of questions acceptably, most buyers stop paying up for the frontier — the same reason you stopped upgrading your phone every year.
Chamath, via @PodcastAlphaXCoding and agents — the fastest-growing slice of token demand, and the one most sensitive to unit cost — are exactly where open models are strongest. That is the segment compounding, not the one being defended.
Axevil CapitalThe story is not open-source killing proprietary margins. It is a geographic shift — and that is a different risk with a different owner. Price it as policy, not as technology.
@glocalinvestorCongress is already probing US firms — Airbnb and Cursor named — for shipping Qwen and Kimi inside their products. Federal procurement bans are the option being discussed most seriously.
@Mayhem4MarketsThe block may come from the other direction first. Beijing has been meeting Alibaba, ByteDance and Zhipu about restricting overseas access to China’s best models. The weights are open until the day they are a strategic asset.
@demian_aiThe security question, answered properly
The strongest argument against Chinese models is not price and not benchmarks — it is trust. Booz Allen Hamilton ran the study everyone cites: roughly 2,800 trials, about 460,000 lines of generated code, prompts written from the personas of a US defence contractor, a Chinese entity and a Russian one.
The headline finding is genuinely alarming. When the prompt said the user worked for the US government, Alibaba’s Qwen3-Coder produced roughly 130% more vulnerabilities than under a neutral persona. DeepSeek and MiniMax showed smaller increases in the same direction. Claude, tested as the control, went the other way and produced more secure code for the same persona.
Now the part that almost never gets quoted, and that changes the trade. Kimi K2.5 recorded the lowest vulnerability score of every model in the study — below the American control. The refusal behaviour is just as uneven:
Booz Allen Hamilton, May 2026: ~2,800 trials, ~460,000 lines of generated code. The spread is the point — these models do not behave alike, and “Chinese model” is not one risk profile.
So “Chinese model” is not a risk profile; it is a passport. MiniMax refuses 80% of sensitive prompts and Kimi refuses 32%, while DeepSeek refuses 8% — less than half of what its compatriots do and only slightly more than Claude’s 2%. Any policy that treats these five as one bucket is going to be wrong about at least three of them. Any investor who treats them as one bucket will be wrong about which one survives a ban.
What the Moonshot pitch prices — and what it doesn’t
Moonshot is raising at $31.5B on more than $300M of ARR — roughly 40–90× revenue, against the ~200× the public market is paying for Zhipu, the same asset class already listed at ~$128B. Same niche, same stage, one is public and one is not: the entry is at something like a 80% discount to the comparable. DeepSeek sits between them at $50B+.
The product argument is also real. Kimi is the only open model that combines open weights, native vision and agent orchestration — GLM-5.2 has no vision at all, and runs a single agent where Kimi parallelises across as many as 300. On the workloads that are actually growing, that combination is the product.
What the deck does not price is the thing we flagged nine days ago. A 40× multiple on API revenue assumes the API stays reachable. If Washington bans federal procurement, or Beijing restricts export of frontier weights, or both, then the revenue line that justifies the entry multiple is the exact line that gets cut. The deck has thirteen pages on why the asset is cheap and none on the one variable that decides whether it is an asset at all.
US vs China: what the multiples actually say
Here is where we have to argue with the people we agree with. The China bull case is usually sold as a value story — overheated American AI, cheap Chinese alternative. On multiples, that is simply false, and we are not going to pretend otherwise to make our own position look better.
On the one metric a value investor would reach for first, the Chinese labs are the expensive ones. Anthropic filed to IPO at $965B on ~$47B ARR — roughly 21×. Zhipu, already listed, trades near 200×. Sources: Axevil Capital (Chinese labs), Anthropic and OpenAI IPO filings, June 2026.
Anthropic filed to go public in June at $965B on roughly $47B of ARR — about 21× revenue. OpenAI filed at $852B, implying around 34×. Meanwhile Zhipu, the listed Chinese comparable, trades near 200×, and Moonshot’s round is being marketed at 40–90×. Foundation-model multiples in the US have actually compressed — from the 60–100× of 2024 down toward the 15–50× band, which is expensive but is not 1999.
So if you buy Chinese AI expecting a discount to the multiple, you are buying the opposite. The Axevil deck is careful about this: it benchmarks Moonshot against Zhipu, not against Anthropic, and calls Anthropic “the ceiling, not the entry.” That framing is doing a lot of work. Anthropic is not the ceiling — it is the cheaper asset.
Then why do we still want the exposure?
Three reasons, none of which is “it’s cheap.”
1. The revenue is not the same kind of revenue. Anthropic committed up to $100B to AWS in exchange for Amazon’s investment, with a parallel arrangement at Google; OpenAI made a comparable commitment to Azure. As Om Malik put it: some portion of every US lab’s revenue is the cloud provider paying the lab to consume the cloud provider’s compute — and some portion of the cloud provider’s AI revenue is the lab paying it back. Both numbers go into both top lines. A 21× multiple on revenue you cannot decompose is not obviously cheaper than 90× on API cash from 200 countries. That is the actual bubble tell, and it is not on any chart.
2. The growth rates are not comparable. Moonshot went from ~$20M to more than $300M of ARR in about six months — a 15×. Anthropic’s multiple is low precisely because its revenue is already large and maturing. You are not comparing two prices; you are comparing a price on a base that has mostly happened against a price on a base that mostly has not.
3. The volume is moving, and volume precedes revenue. Nearly half of OpenRouter’s tokens are already Chinese. Six Chinese models rank above Claude by volume. That flow monetises later, and it monetises into a business with a fraction of the compute bill.
Anthropic is worth 31× Moonshot. For Anthropic to double, someone must underwrite another trillion dollars of AI equity. For Moonshot to double, it needs to reach Zhipu’s current price — a company in the same niche, at the same stage, already public.
Our position, stated plainly
We want exposure to independent Chinese AI labs, and we are prepared to pay a higher revenue multiple to get it — because the growth is earlier, the revenue is cleaner, and the volume has already moved. We think the US AI complex is where the crowding is: two labs carrying $1.8 trillion of private equity value between them, on revenue that is partly a round-trip with their own cloud investors, at a moment when the customers those revenues depend on are actively migrating to cheaper models to cut their bills.
We are not saying Chinese AI is cheap. It is not. We are saying the American price embeds an assumption — that the frontier stays worth 7× the alternative — which the routing data says is already breaking.
!The one thing that would flip us: the geo-block. A federal procurement ban, or a Beijing export restriction on frontier weights, does not dent the thesis — it deletes it. This is not a risk we can hedge, only one we can size for. Anyone telling you they can own this trade without owning that risk has not read the last month of news.
The TT desk thoughts
We were right on the mechanism and we want to be honest about what we got wrong. We said a “4% quality gap.” On SWE-Bench Pro the real gap between Opus and the best open model is 7.1 points, and against Kimi specifically it is 10.6. The gap is wider than we said — and it did not matter, because the market bought on price anyway. That is worth sitting with: the commoditisation did not need parity to happen. It only needed “good enough, at a fifth of the cost,” and it got that a year ago.
The call stands: do not own the model layer. Price per token has fallen ~85% for the same output quality and there is no floor in sight, because the marginal supplier is a lab that gives the weights away to win distribution. Owning a model business here is owning a commodity producer that is being subsidised by its competitors’ strategy. The value accrues where it always does when the input commoditises — to distribution, proprietary data, and the rails that route and verify. OpenRouter itself is the tell: the most valuable position in this chart is not any of the models on it, it is the switchboard that made them substitutable.
The new call, and the one we are actually acting on: the geo-block is now the dominant variable, and it is not priced. Nine days ago it was a footnote in our own post. Today Congress is probing named US companies for using Kimi and Qwen, Booz Allen has handed the hawks a citable study, and Beijing is reportedly weighing an export restriction of its own. Either side moving turns a cost advantage into a stranded asset overnight. Concretely, we would rather own the exposure that survives both outcomes — inference infrastructure, routing and eval layers, and the enterprises whose margins just structurally improved — than a bet that a specific Chinese lab’s API is still callable from Delaware in eighteen months.
Where we are looking for the entry: the private markets. The desk is actively working on access to independent Chinese AI labs — Moonshot, DeepSeek, MiniMax — through pre-IPO and secondary channels, because that is the only place the exposure exists on the terms we want it. Zhipu already listed and repriced to ~200×; the entry we are looking for is the one that happens before that repricing, in the same asset class. The public route into this theme — a US-listed AI basket — buys you the crowded side of the trade, not this one.
Three things to watch, in order of how much they would move us: whether a federal procurement ban is actually written rather than floated; whether Beijing restricts overseas weight access (this is the one nobody is positioned for); and whether the Opus-class gap widens back out, which is the only path by which the frontier earns its price again. If the gap widens, the model layer stops being a commodity and this whole note is wrong. We do not think it will — but that is the trade we are on the other side of, and you should know it.
Keep reading
Do AI margins move from the labs to the infra? · The ASI transition: naive, or naive squared? · AI × crypto — the full framework